利用階層式類神經網路實現車牌辨識系統

本論文提出利用階層式類神經網路實現車牌辨識系統，系統功能區分：影像前處理、車牌區域定位、字元分割以及字元辨識四大模組。其中影像車牌區域定位是採用連通標記法的方法，輔以區域分析將正確的車牌影像擷取出來。字元分割模組以投影量分割法為基礎，並使用雷登轉換來校正車牌的傾斜角度。而字元辨識模組，以階層式架構來建構辨識系統，其中階層式倒傳遞類神經網路扮演了字元辨識的角色。辨識器將字元依序讀入並先做數字的辨識若未達設定門檻值，則進行第二階段英文字元之辨識。由實驗結果發現所提出的辨識系統可高達到 98.9% 的車牌影像定位與擷取，而字元辨識率達99.7 %，而正確車牌辨識率可達到 98.4% ；由實驗結果驗證所提出的方法應用在車牌影像辨識上有很好的辨識效果。

關鍵字

影像處理；倒傳遞類神經網路；車牌辨識

並列摘要

This thesis proposes a license plate recognition system with image processing equipped with the hierarchical neural network architecture. The overall system can be divided into four modules: image preprocessing, zone localizing for license plae, character pattern segmentation, and character recognition. The zone localizing process uses the connected component method associated with the zone characteristic analysis and it can indicate the region of the license plate on entire car images. The character pattern segmentation uses the Radon transform to detect the tilt angle and the vertical projections for charterer segmentation from the captured license plate region. The character recognition module is implemented by the hierarchical back propgation neural network (BPN) architecture, where digit and letter patterns are separately recognized. In the digits’ recognition, a given matching threshold is given for scoring and labeling. The other unlabeled patterns then are sent to the letters’ recognition stage. The experimental results show that the proposed recognition system can achieve 98.9% correct rate; where the success rate of character recognition is 99.7% and the license plate locatlizing can reach the correction 98.4%. Thus, the proposed recognition system can be demostrated to be efficient and feasible in this study.